Meituan LongCat-2.5

Meituan's new 1.6-trillion-parameter multimodal model for long agentic coding tasks is live in its demo and API, with weights promised soon.

LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token, built on LongCat-2.0 with native image understanding (images are turned into discrete tokens) and more long-horizon agent and coding training. Meituan says it was trained and served on domestic Chinese AI accelerators, including a 4-bit inference setup. Its own benchmark table compares it with GPT-5.6-Sol and Claude Opus 5; those are Meituan's numbers. The online demo and API are open now; Meituan says GitHub and Hugging Face weights will follow, and as of Nov 2 its Hugging Face page still tops out at LongCat-2.0. Checked Nov 3 (about 14:00 UTC): Meituan's Hugging Face organization still has no LongCat-2.5 repository.

Status
Released
Date
Nov 2, 2026 (blog dated Nov 1)
Maker
Meituan
Type
AI model
Categories
AI

Sources

Launch dates can change. Check the sources for the latest. Last reviewed Nov 10, 2026.

Related rumors